Research On Radar Signal Recognition Based On Improved UNet3+ And Undersampling

The identification of radiation source signal plays a very important role in the field of electronic countermeasures reconnaissance.It is the key basis for command decision and judgment to obtain various information in the battlefield by identifying radar signal.The electromagnetic environment of the early battlefield is relatively simple,and the traditional method mainly extracts the signal characteristics of the radiation source manually,and then compares with the radar database to identify the radar signal,which has good identification ability.However,with the continuous innovation of radar technology,various radars based on new technology are constantly applied in practice,and some signals are difficult to detect and receive,and the electromagnetic environment is increasingly complex,the identification speed and accuracy of this method can not meet the identification requirements gradually,and the reliability is greatly reduced.Therefore,an effective signal recognition method is needed in the complex electromagnetic environment to meet the increasing wartime needs.Deep learning is a data-driven algorithm that can extract more effective information from radiation source signals and has excellent performance.In this paper,the method of deep learning is introduced into radiation source signal recognition.The research contents are as follows:(1)Several basic neural network models are briefly introduced,and their main characteristics and application scope are analyzed.At the same time,three classical data imbalance processing methods are analyzed,and their principles,advantages and disadvantages are analyzed,which provides theoretical foundation for the follow-up work of the paper.(2)In view of the problems of traditional radiation source signal recognition method being time-consuming,requiring manual feature extraction and poor signal recognition ability for low SNR,this chapter proposes a recognition method based on improved Unet3+ network.The input of the model is sequence data,and its feature extraction is carried out in the network without manual feature extraction.The attention mechanism is introduced into the network model to improve the training speed and recognition accuracy.Through experiments on 8 kinds of simulation signals,the results show that the improved model has a short training time and can effectively identify signals with low SNR.(3)Aiming at the problem of biased recognition ability of neural networks trained with categorically unbalanced data sets,this paper presents an improved undersampling recognition method,this method is the first to use most of the class of training samples of the network to get a classifier,and then for continuous from the distillation of classifiers to improve its generalization ability,Finally,undersampling method is used to obtain a balanced data set to train the network.And it turns out,for the three minority class signals,the recognition accuracy of the improved undersampling model is 92.42%,89.71% and 92.42%,which better improves the network bias,and the recognition accuracy of all kinds of signals is more balanced.Finally,the network model can not only solve the problem of network bias,but also avoid the significant decrease of signal recognition accuracy of most classes.